Proposal: Reduce deployment cost — make AgentFlow a zero-config OpenClaw native plugin

[Tweakzx]

Problem

Current AgentFlow deployment requires:

1. pip install -e . — Python environment setup
2. agentflow init --db ./data/agentflow.db — Manual DB initialization
3. agentflow serve --db ./data/agentflow.db --port 8787 — A persistent Python process
4. The OpenClaw plugin shells out to CLI (execFile("python3", ["-m", "agentflow.cli", ...])) for every operation

This is heavy for users who just want "agent task tracking + visualization". The Python service is a separate lifecycle to manage (start/stop/crash/restart), and the CLI bridge adds latency and fragility.

Proposal: Pure TypeScript OpenClaw Plugin

Rewrite the core as a pure TypeScript OpenClaw plugin that embeds better-sqlite3 (or better-sqlite3 bundled) instead of shelling out to Python.

Architecture

plugins/agentflow/
├── openclaw.plugin.json     # Plugin manifest
├── package.json             # Dependencies: better-sqlite3
├── index.ts                 # Plugin entry (register tools/commands/routes)
├── store.ts                 # SQLite wrapper (same schema, pure TS)
├── broker.ts                # In-process event broker + SSE
├── api.ts                   # REST API handlers (embedded in gateway)
└── console/
    └── index.html           # Static dashboard (served from plugin)

What Changes

Current	Proposed
Python service (separate process)	Embedded in OpenClaw gateway (same process)
pip install -e .	npm install (or bundled with plugin)
agentflow init	Auto-init on first use
agentflow serve --port 8787	Served on gateway port (e.g., /agentflow/)
CLI bridge (execFile python3)	Direct function calls (zero IPC)
Separate port 8787	Gateway port (already running)
Python tests	TypeScript tests

What Stays

• Same SQLite schema — projects, tasks, runs, triggers, gate_profiles, etc.
• Same REST API — /api/tasks, /api/task/<id>/move, /api/flow, etc.
• Same Web Console — Same HTML/CSS/JS, served as static file
• Same OpenClaw plugin surface — tools, commands, HTTP routes

Concrete Benefits

1. Zero-config install: openclaw plugin install agentflow → done
2. No separate process: DB operations happen in-process, no IPC overhead
3. Auto-init: First tool call creates the DB automatically
4. Gateway-embedded UI: Dashboard at http://gateway:3000/agentflow/ — same auth, same port
5. Better reliability: No process management, no Python path issues, no PYTHONPATH hacks
6. Type safety: Plugin, store, and API all in TypeScript

Migration Path

• Phase 1: Reimplement store.ts with better-sqlite3 (same schema)
• Phase 2: Reimplement API handlers as OpenClaw HTTP routes
• Phase 3: Bundle the HTML console as a static asset
• Phase 4: Deprecate the Python CLI (keep for standalone use)

Why Not Fork

The Python CLI is still useful for:

• Standalone use without OpenClaw
• Scripting and automation
• Non-OpenClaw environments

But the OpenClaw plugin path should not depend on it.

Open Questions

1. Should the Python CLI remain as a separate package, or live alongside the TS plugin?
2. Should the DB path be configurable via plugin config, or default to ~/.openclaw/agentflow.db?
3. Should we use better-sqlite3 (native, fast) or sql.js (WASM, no native deps)?

Related Issues

• Feature: Add POST /api/tasks endpoint for programmatic task creation #5 (POST /api/tasks) — easier to implement in TS
• Enhancement: OpenClaw plugin — expose full task CRUD via tools/commands #7 (Plugin CRUD tools) — direct function calls, no CLI bridge
• Enhancement: OpenClaw plugin tools return raw CLI stdout instead of structured data #18 (Structured output) — tools return JSON natively
• Bug: Store.connect() creates a new connection per call — no connection pooling #19 (Connection pooling) — better-sqlite3 handles this natively